CO2 sequestration in saline aquifers is a crucial component of carbon capture, storage, and utilization (CCUS) technology. The subsurface fluid flow of CO2 and brine in porous media involves the coupling of multiple physics fields, featuring complex nonlinear partial differential equations (PDEs). The prevalent approach for studying subsurface fluid flow is to discretize PDEs in spatial and temporal dimensions and solve them numerically. Here, this work proposed a Gradient-guided Convolutional AutoEncoder (GCAE), where the gradient differential operator is incorporated as physical prior knowledge into the loss function of the neural network. The physical prior knowledge guides the training process of the neural networks, enhancing their physical interpretability compared with the purely data-driven Convolutional AutoEncoder (CAE). This work applied GCAE to the CO2 sequestration in the homogeneous formation, the heterogeneous formation, as well as the heterogeneous formation with different well placements to demonstrate the improvement in prediction accuracy, data stability, and generalization capability compared with the CAE approach.

Gradient-guided Convolutional AutoEncoder for predicting CO2 storage in saline aquifers with multiple geological scenarios and well placements